Improved Reservoir Management Through Optimal Control and Continuous Model Updating

Abstract

Abstract There is a potential for large improvements in reservoir management by using optimization and model updating techniques in a closed-loop fashion. Here we demonstrate how the combination of the ensemble Kalman filter technique for continuous model updating with an automated adjoint-based water flood optimization algorithm leads to significant improvements in net present value (NPV) of the water flooding process. Using the ensemble Kalman filter, both static parameters (permeabilities) and dynamic variables (pressures and saturations) are updated in the reservoir model as new production measurements become available. Other properties are assumed known in advance. At the start of the production process, in the absence of information on the permeability distribution, an optimal control strategy based on a homogeneous reservoir is used. Subsequently production data are at regular intervals assimilated with the ensemble Kalman filter, resulting in an updated estimate of the reservoir pressures, saturations and permeability field. Based on these updated parameters an optimal water flooding strategy is determined for the remainder of the production process. This process of model updating and optimization is continued over the life of the reservoir. The methodology is applied to two synthetic examples, enabling comparison with traditional production strategies. Significant improvement in NPV, acceleration of oil production, cumulative oil recovery, and reduction of water production were realized. Results were close to those obtained with water flood optimization based on an a-priori known reservoir description. For one example the improvement in cumulative oil recovery is about 44 %, which is quite close to the improvement obtained for an a-priori known reservoir description. Introduction There exists a potential for large improvements in reservoir management by using the measurement and control opportunities nowadays available in the oil industry. One way to fully exploit the control valves is to optimize their settings in order to maximize the net present value (NPV) of the production process. Various studies conducted on numerical reservoir models showed that significant improvements in the production process may be feasible by dynamically controlling the valves. The exact scope varies with geological features, well operating constraints, and well architecture[R1]. Furthermore, the scope depends on whether the objective is to optimize injection and production rates on the short term, or to optimize the production process over the producing life of the reservoir. A literature overview is given in Ref 1. A limitation of these studies is that they are conducted on reservoir models with all properties known a-priori. In reality, the valve settings have to be computed using the available information there is on the reservoir. Before production starts, a reservoir model has to be built on data from seismic, well tests, core samples, etc. Because of geological uncertainties such models are usually only a very crude approximation of reality, and model parameters, such as permeabilities and porosities are only known with a large degree of uncertainty. Therefore the predictive value of such reservoir models is limited and tends to deteriorate over time. To improve its predictive capacity the model may be adapted such that predicted results approach measured production data. In the oil industry this process is general referred to as ‘history matching’. Because production data are used for model updating, the management of a reservoir based on history-matched models could be considered a closed-loop process. Unfortunately, traditional history matching suffers from a number of drawbacks:It is usually only performed on a campaign basis, typically after periods of years.The matching techniques are usually ad-hoc and involve manual adjustment of model parameters, instead of systematic parameter updating.Uncertainties in the state variables, model parameters and measured data are usually not explicitly taken into account.The resulting history-matched models often violate essential geological constraints.